The Comprehensive Guide to Product Lifecycle Management

Part 1 of 4, from the Mendix Community Netherlands Round Table in Amersfoort, June 2nd, 2026. Four questions came out of the evening: what is left for the consultant to do, whether we can trust an agent to do it, whether the platform still earns the choice, and who is still buying.

Roughly twenty Mendix consultants spent an evening arguing about one thing, even when they thought they were arguing about something else. Not whether the agent can build the app. We mostly agree it will. The argument was about what that leaves for us to do.

This was the Round Table on June 2nd, in Amersfoort. Three tables: Alpha, Delta and Foxtrot. A bank of 34 statements about agentic AI, a live app to vote in, and a generous hour to fight it out. Sixteen of them voted, 42 votes in all. What I was glad to see afterwards, reading it all back, was that nobody at the tables actually panicked about being replaced. The mood was not “are we finished?” It was “so what is the job now?”

So what were we actually arguing about?

‍

Understand, analyze, build

One table put a clean frame on it. Software work has three mental stages: understand, analyze, build. The path from a vague request to a working thing.

Here is the uncomfortable part. The agent is getting very good at the build. The typing. Almost everything worth saying about the next few years sits in what that does, and does not, do to the other two.

A story from Alpha made it concrete. A consultant had been on a project building a customer portal, moving a company off paper forms and into a Mendix app. The product owner had spent two weeks writing user stories, clean ones; every form turned into a neat ticket. Then someone asked the question that should have come first: what did we actually sell here, do we automate paper or do we automate processes? We automate processes. The stories were wrong. Two weeks of careful work, pointed at the wrong thing, and all of it had to be written again.

Now run that through an agent. You get the same great stories, he said, but not the correct thing. Faster, cleaner, and just as wrong. Or, as the table put it: AI is not making a mistake. It is doing exactly what you told it to.

That is the whole argument in one anecdote. The agent multiplies your framing. If the framing is wrong, it multiplies that too, at speed. Understanding the real problem and deciding what is actually worth building are not the easy parts you rush through on the way to the code. They are the part that decides whether the code was worth writing.

‍

Does that mean only the seniors survive?

You can read all of that as bad news for juniors. If the value is judgment, hand the work to the people who already have it and let the agent cover the rest. The room did not buy it. Asked whether teams now only need senior orchestrators and no juniors, 80% said no. The reasoning was simple. Judgment does not arrive with the job title. You build it by doing the work, including the parts a machine can now do for you. A table that quietly stops growing juniors is a table that runs out of seniors in a few years, with nobody left who learned how the thing actually fits together.

The shape people kept reaching for was not “expert” but T-shaped. A project with five T-shaped people beats a project with five experts, as one consultant put it: people who challenge the business and challenge the testing, instead of “I just want to do pure development.” The agent does not remove the need for that. If anything, it raises it.

‍

So what did we keep calling the value?

Strip out the noise, and the same answer surfaced, in different rooms and different words. The value was never really the speed.

What they named instead was the unglamorous, durable stuff. The model two people can read together. The conversation between business and IT that happens around it. Even if the AI generates the whole thing, one consultant said, it is still good that you can fall back on a readable application and go through it together. When a production issue lands on your desk at the wrong hour, you want a flow you can read, even if an agent wrote it.

That readable model is part of what is still ours: holding the conversation, owning what ships. Whether it is also the lasting case for Mendix as a platform, the question that split the room hardest, is a fight for another day.

‍

So far this sounds comfortable. It isn’t.

One thing should keep you honest, and it is not the one you would expect. The new breadth carries a new risk. Being T-shaped is good until it is too much. Ask an AI for a range date picker, and it will happily find one on GitHub and install it, along with whatever someone hid inside it. You have an info-stealer on your website without knowing it, as one consultant put it. You are T-shaped a little too much. The agent will reach for tools you never vetted, and the responsibility for what it reached for is still yours.

Which is why the cleanest idea of the night was not about trusting the agent at all. It was about boundaries: give it read-only access, do not hand it the tool to do the dangerous thing, and put a human on the actions that are hard to undo. Gate the irreversible, not the agent. And, said on the record by the host of that table: the gate is for now, not forever. It will move to three, without a doubt. Trust gets earned, gates come off, and today’s careful consensus has an expiry date.

‍

So what does the consultant become?

Less of a builder, for sure. The agent is taking the build, and it is welcome to it. What is left is the part that was always the actual job, the part we sometimes hid behind the typing: working out which app is worth building, holding the conversation between the people who have the problem and the system meant to solve it, deciding what is safe to automate, and being able to read and stand behind what ships.

The agent can write the app. It still cannot tell you which app is worth writing, and it will not be in the room when the wrong one goes live. That, it turns out, was always the job.

That is the first of the four questions the evening threw up. The next one sits directly underneath it: if the agent is doing the building, can we trust it to? More on that next time.

We got through a handful of the 34 statements. The rest are still in the bank. What if we continue the conversation online? Let me know if you would be interested, and I will figure out a format to do so.

‍

Originally published here.

Every AI conversation in industry right now circles the same uncomfortable question: we have invested, we have piloted, we have demoed, so where is the value? At Realize LIVE 2026, Siemens gave its answer. It's called Intelligence Center X (ICX) and for everyone building on Mendix, it's the moment the platform steps into a new role: from low-code development platform to the agentic heart of Siemens' industrial AI strategy.

‍

From vision to product: Realize LIVE 2026

Siemens CEO Roland Busch has delivered one consistent message on every major stage in 2026: industrial AI is leaving the lab. Siemens, in his words, is delivering "AI-native capabilities, intelligence embedded end-to-end across design, engineering and operations", not AI as a bolt-on assistant, but intelligence woven into how products are designed, built, and operated.

At Realize LIVE Americas 2026 in Detroit (June 1–4), that message became a product. In front of roughly 3,000 users and partners, Tony Hemmelgarn, President and CEO of Siemens Digital Industries Software, announced Intelligence Center X: industrial AI orchestration software designed to turn AI from isolated experimentation into scalable, governed business impact with the Mendix platform at its core.

A few weeks later, at Realize LIVE EMEA in Amsterdam, the topic was impossible to escape. Whether the session was about Teamcenter, Opcenter, simulation, or Mendix itself, every conversation kept returning to the same question: how do we make our processes intelligent, safe, at scale, and with proof that it works?

‍

The real problem: four gaps between AI ambition and AI value

Why a new "Center" alongside Teamcenter X, Simcenter X, and Opcenter X? Because across industries, AI initiatives keep stalling in the same four places.

‍

Gap #1 – Software agents are vulnerable

Agentic AI is powerful precisely because it acts autonomously and that is also its risk. An agent that reads enterprise data and triggers actions can be misled by stale data, manipulated inputs, or ambiguous instructions. Hemmelgarn quoted one customer CEO in Detroit: "The last thing I need in my organization is for AI to go grab and lock on to a SharePoint location that's 20 years old." An agent confidently acting on untrusted data isn't automation; it's a liability.

‍

Gap #2 – Shadow AI and limited oversight

Teams don't wait for IT. Engineers wire up their own copilots, departments subscribe to point solutions, and someone in quality is already running a model nobody approved. The shadow IT problem of the early cloud era is repeating itself with AI,  faster and with higher stakes. The result: ungoverned agents that IT can neither see, secure, nor switch off.

‍

Gap #3 – Workflows and agents have no audit trail

When a human approves a change, PLM captures who, when, and why. When an AI agent recommends absorbing a cost instead of redesigning a part, who signed off? On what data? Under which policy? Most AI experiments cannot answer these questions, and for regulated industries that is disqualifying.

‍

Gap #4 – Fragmented visibility across the portfolio

Even organizations with successful AI use cases usually can't see them as a whole. One model in manufacturing, one agent in the supply chain, a copilot in engineering, each with its own data connection, owner, and definition of success. There is no single place to see which agents exist, what they may do, and what they deliver.

‍

How Intelligence Center X closes the gaps: four parts, one trust layer

Intelligence Center X treats these gaps as one architectural problem, bringing four capabilities together on a single governed foundation and if you know Mendix, you already know half of the stack.

‍

1. Knowledge graph with Graph Studio

Graph Studio builds the enterprise knowledge graph: it connects data from engineering, manufacturing, supply chain, and service, and sets up an ontology between them, a living semantic model that gives every data point meaning and relationships. Out-of-the-box industrial ontologies accelerate the start. This context layer is what closes Gap #1, agents reason over connected, current, trusted data instead of stale copies in data lakes.

2. Machine learning with AI Studio

AI Studio is where data scientists and engineers build, train, and operate machine learning models, grounded directly in the data contextualized by the knowledge graph. Instead of exporting data into yet another isolated ML environment, models connect natively to the governed graph, so predictions inherit the same context and lineage as the data they were trained on.

3. Agentic development with Mendix

This is where intelligence becomes actionable and where Mendix shines. Teams build the applications and AI agents that put insights to work: human-in-the-loop apps on the shop floor, autonomous agents for routine decisions, and everything in between. Agents are modeled, versioned, and deployed like any other Mendix artifact visible to IT and governed from day one. That is the structural answer to shadow AI (Gap #2), a sanctioned, productive place to build agents beats a ban every time. And for existing Mendix landscapes, integration into the ICX tooling is refreshingly non-dramatic. Your apps, modules, and DevOps pipelines connect to Graph Studio context and AI Studio models through standard connectors and the Model Context Protocol (MCP), not a rip-and-replace migration.

‍

4. Process orchestration with Mendix Workflows

Individual agents create tasks, orchestrated agents create value. Mendix Workflows coordinates people and agents in end-to-end business processes: an agent detects an anomaly, a workflow routes the finding, a human approves the action, an agent executes it. Every step is part of one traceable process definition, which is exactly what turns AI from a side experiment into an operating model.

The foundation: an enterprise trust layer. Everything above runs on a shared trust layer that answers Gaps #3 and #4 and makes the system enterprise-ready. It provides:

• guardrails defining what agents and humans may do and when a human must step in
• traceability through end-to-end logging of every action, so every decision has an owner, a timestamp, and a data lineage
• security built on a pen-tested platform with a hardened runtime and real-time anomaly detection
• ready-to-use governance with policies, integrated DevOps, Identity & Access Management, and the Control Center as a single pane of glass across the entire portfolio

‍

Proof it works: Vivix

None of this is theoretical. Vivix Vidros Planos, Brazil's leading flat glass manufacturer, built its flagship "Smart Furnace Monitoring" initiative, watching over a $120 million furnace, on nearly 30 Mendix applications connecting OT and IT data across SAP S/4HANA, Siemens Industrial Edge, and Snowflake. On top sits their AI-powered Virtual Engineer: an assistant that gathers product, production and process data, and gives quality and production teams tailored suggestions, so they act proactively instead of reactively. It is built on Mendix with Amazon Bedrock and Claude from Anthropic.

The results ended the pilot-purgatory debate: an 85 percent reduction in production issue resolution time, 6,000 hours of manual work recaptured in a single year, customer complaint resolution compressed from five days to under one, and up to 4x faster resolution in quality-related investigations, recognized with a Siemens Techcellence Award and an AWS GenAI Gamechanger award. Mendix is the layer that made the intelligence actionable, turning connected data and models into applications people use every day.

Mendix 11.12: the first LTS release ready for agentic production

Announced around Realize LIVE EMEA, Mendix 11.12 turns the agentic promise into something you can put into production, because it is the first Long-Term Support (LTS) version of Mendix 11, the first LTS since 10.24. That matters more than any single feature.

Mendix best practice is clear: production applications belong on LTS versions, where you get a stable, long-supported foundation instead of chasing monthly releases. Until now, teams that wanted Mendix's agentic capabilities had to build on moving ground. With 11.12, agentic development is LTS-grade for the first time, meaning you can build agents and take them to production following the same release discipline you already apply to your business-critical apps. The release also delivers Agents Kit 2.0 with the Agent Editor in Studio Pro, built-in MCP server and client components, big performance gains (25–40% faster project loads, up to 6x faster error checking, up to 8x faster local deployments), and embedded Mendix workflows inside Teamcenter Active Workspace.

‍

Mendix 11.12 highlight: Maia

The single biggest reason 11.12 feels like a step change is Maia, Mendix's AI assistant, now woven through the development lifecycle.

Describe the solution, and Maia helps build it. You start from intent rather than a blank canvas: you explain what the solution should do (the process, the data, the outcome) and Maia generates the structure to match. Maia Plan turns that into epics and user stories (flowing straight into your Jira backlog), and Maia Make begins implementing the scoped work in Studio Pro. You direct; Maia drafts.

Maia's value isn't a one-shot code dump. It works alongside you throughout the build, actively supporting your own development, suggesting the next step, filling in the repetitive parts, catching gaps, while you stay in control of the design. It behaves like a capable pair-programmer who never tires of the boilerplate.

Additionally, low-code makes the output reviewable. What Maia produces is a Mendix model, not thousands of lines of raw code, and a model is far easier to read, review, and reason about than hand-written code. You see the microflow, the page, the workflow, the data model at a glance and immediately judge whether it's right. This is also the core benefit of low-code in the AI era: AI allows companies to move fast, while making it possible for a human to meaningfully verify the result. Speed without a black box.

A personal note on live demos. It changed my own experience. In the past, a live demo meant coding on the fly, explaining, and hoping nothing breaks, all at once. Now I prepare user stories ahead of time and simply run them during the demo, and walk the audience through each task and its result calmly. The demo went from a tightrope act to a conversation, from performing under pressure to explaining with confidence.

‍

Mendix as an MCP server for other tools

Maia is the built-in path, but not the only one. Because 11.12 ships with MCP support, a Mendix application can act as an MCP server, exposing its logic, data, and workflows as tools that other AI clients can call. You are not locked into a single assistant: your governed Mendix capabilities can be consumed by another agent or tool of your choice, while still running behind the ICX trust layer. This is precicely the difference between an AI feature and an open, interoperable AI platform.

‍

Integrating and orchestrating AI agents with Mendix

The same openness runs in the other direction. Mendix doesn't just expose tools, it consumes them, using its MCP client to bring external AI agents and services into your applications. That lets you combine best-of-breed agents from different providers, and then do the part that really matters: orchestrate them with Mendix Workflows. Rather than a loose collection of agents each acting on its own, you get coordinated, human-in-the-loop processes where Mendix decides which agent runs when, hands off between agents and people, and keeps every action inside the governed, auditable trust layer. Mendix becomes the conductor of your hybrid workforce, not just one more instrument in it.

‍

Personal takeaway

Intelligence Center X is, to me, a genuinely strong concept: it connects the data of very different systems and makes it usable for a whole range of scenarios, AI foremost among them. And within that concept, Mendix is the key player. The combination of agentic development and low-code is what makes the difference. It lets you build AI-driven solutions fast, but also review, govern, and trust what you've built. That balance of speed and reliability is exactly what enterprise software needs and it makes Mendix a strong partner for reliable, efficient software solutions in the age of industrial AI.

Much of our daily work consists of gathering information, searching through resources, rewriting the same ideas in different formats, or summarizing discussions. This kind of grunt work naturally makes people look for ways to save time, reduce repetition, and reach a useful outcome with less friction. And agentic AI presents itself as a powerful shortcut, capable of turning a rough thought into a polished email, a meeting into a concise summary, or a manual workflow into an automated step.

For many teams, that is precisely the trap: that AI chatbots and copilots are capable of the same kind of conceptual work people do instinctively when taking a vague goal, breaking it into meaningful parts, and moving it toward a useful outcome. And they, eventually, become lazy in the wrong way.

AI adoption starts with realistic expectations

The public promise around AI adoption suggests something close to a general-purpose thinking partner. A system that can understand goals, interpret context, and work through a problem much like an experienced colleague would. But that couldn’t be further from the truth.

AI is a predictor, not an understanding system

At its core, AI predicts likely continuations. It does not understand your business model, your operating constraints, or the relationships between the moving parts inside your organization in the way a person does. It does not know why one exception matters more than another, why a process exists in its current form, or what tradeoffs sit behind a decision unless those things are made explicit.

As a result, AI can be fast, but still shallow. It may generate something that looks complete while missing the mechanics that actually matter to the business.

AI can sound professional and still be wrong

One of the reasons AI is so persuasive is that it presents information with confidence. It writes clearly, and structures arguments well, but fluent output should not be mistaken for expertise.

An experienced professional in any domain will typically outperform AI where nuance, context, and consequence matter. They can spot what is missing, question what does not fit, and recognize when an answer is technically plausible but practically wrong. AI, by contrast, can hallucinate, or flatten important distinctions.

More context does not automatically make AI better

The model does not understand information the way a person does. It does not inherently know which detail is strategically important, which relationship between inputs actually drives the outcome, or which exception should outweigh the broader pattern. When too much context is added without clear structure, prioritization, or framing, important information gets diluted by secondary or weak signals, and the output reflects surface-level correlations.

This is why strategic selection still remains a human responsibility.

The human role in AI strategy and implementation

A properly structured and governed AI strategy is what turns agentic AI from a shortcut into a system. But that structure does not come from the model itself. It comes from the person using it. The one that determines what information matters, what can be ignored, what the actual goal is, and which constraints should shape the output. AI does not make those decisions well on its own, at least not reliably in its current state.

For this reason, every human-agent exchange should be guided by five essential elements:

1. A scope with boundaries

AI performs best when the task is defined clearly enough that the system knows what it is being asked to do and what sits outside its role. Without boundaries, AI tends to default to broad, generic responses that may look complete but lack operational relevance.

A system asked to support a service workflow, for example, needs to know whether it is drafting a response, classifying an issue, extracting information, or recommending a next step.

2. A defined process

AI is most effective when it supports a process rather than replacing one that has never been properly defined. If the workflow itself is unclear, inconsistent, or heavily dependent on undocumented workarounds, AI will reflect that ambiguity, not resolve it.

This is often where organizations overestimate the technology. They assume the model can compensate for process gaps, while in practice it usually amplifies them.

3. A definition of quality

If no one defines what “good” looks like, AI cannot reliably produce it. Quality has to be made explicit. That means deciding what level of accuracy is acceptable, what kinds of mistakes matter, what should trigger review, and where the cost of error is too high for guesswork.

In customer communication, “good” may mean clarity and consistency. In regulated documentation, it may mean traceability and compliance. In an internal support workflow, it may mean speed with a human checking exceptions. AI cannot infer those standards reliably on its own.

4. Curated context

Context is valuable only when it is relevant, structured, and timed correctly. That means selecting the inputs that genuinely influence the task, excluding what does not, and making the relationships between inputs as legible as possible.

In a manufacturing setting, that might mean prioritizing machine status, order constraints, and maintenance windows over general historical information, while in a commercial process, grounding the system in account history, product rules, and current workflow stage.

5. Controlled rollout

AI should not move from a promising output to full autonomy in a single step. Before a system is trusted in real operations, teams need to test it in practice, compare its output against human judgment, and define where review remains necessary.

Let AI run beside people first, supporting the workflow, exposing patterns, proving its reliability, and monitoring performance over time. Teams need to see where the system performs well, where it fails, what kinds of errors recur, and when changes in context or model behavior require adjustment. That is how trust is built.

A practical framework for AI implementation in business

Many AI implementation efforts stall because organizations move to agents before they are ready. At CLEVR, we often see this happen when the process is not yet clearly defined and trust has not yet been built into the workflow. That is why we have developed a six-step methodology for organizations preparing for the agentic future.

It is the same framework we use in our advisory process to assess AI maturity, identify structural gaps, and define the right next step on the path toward more advanced agentic systems:

1. Understand the work

Look at how the process actually runs in practice, including manual workarounds, exceptions, and informal steps that never appear on an org chart.

2. Find the real opportunities

Identify where AI can create meaningful leverage without introducing unnecessary risk (e.g. processes that include repeated effort, predictable bottlenecks, valuable time being lost on manual work, etc.)

3. Design deliberately

Decide what the AI is supposed to do, what information it needs, and what should remain within human scope.

4. Define what “good” means

Determine what acceptable performance looks like, what kinds of mistakes matter, which outputs require review, and where the cost of error is too high for approximation.

5. Run beside people first

That makes it possible to compare outputs, identify failure patterns, and learn where human review still adds the most value.

6. Monitor and improve

The strongest AI systems improve because they are observed, adjusted, and scaled from proven use cases rather than assumed to be finished from day one.

Make your AI strategy the real leverage

AI can create a unique competitive advantage, but only when it is applied with structure, context, and a clear definition of success. That is especially true for organizations looking beyond prompts and toward agents, automation, and more autonomous systems. With 30+ years of experience in digital transformation, and a strong portfolio in AI agents, AI solutions, and automation, CLEVR helps organizations take those first steps with clarity.

From understanding where AI can create value, to designing the right use case, and building a roadmap toward more advanced agentic systems, we help turn early ambition into practical progress. One step, one use case, and one trusted outcome at a time.

‍